Journal of Health Economics 31 (2012) 271–284

Contents lists available at SciVerse ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Financial incentives for maternal health: Impact of a national programme in

Nepal
Timothy Powell-Jackson ∗ , Kara Hanson
London School of Hygiene and Tropical Medicine, Health Economics and Systems Analysis Group, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: Financial incentives are increasingly being advocated as an effective means to influence health-related
Received 3 November 2010 behaviours. There is, however, limited evidence on whether they work in low-income countries, particu-
Received in revised form 25 August 2011 larly when implemented at scale. This paper explores the impact of a national programme in Nepal that
Accepted 28 October 2011
provides cash incentives to women conditional on them giving birth in a health facility. Using propensity
Available online 3 November 2011
score matching methods, we find that the programme had a positive, albeit modest, effect on the utili-
sation of maternity services. Women who had heard of the SDIP before childbirth were 4.2 percentage
JEL classification:
points (17 percent) more likely to deliver with a skilled attendant. The treatment effect is positively asso-
I10
I18
ciated with the size of the financial package offered by the programme and the quality of care in facilities.
Despite the positive effect on those exposed to the SDIP, low coverage of the programme suggests that
Keywords: few women actually benefited in the first few years.
Impact evaluation © 2011 Elsevier B.V. All rights reserved.
Financial incentives
Demand for health care
Maternal health
Nepal

1. Introduction
Access to priority health services in low-income countries
remains vastly inadequate. Nowhere is this more obvious than in
maternal health. According to a widely cited paper by Campbell
and Graham (2006), a strategy in which women give birth in pri-
mary care institutions with effective referral is key to improving
maternal health. Yet, improvements in the coverage of professional
care at childbirth has stagnated in Sub-Saharan Africa and South
Asia over the past decade, in part, because the provision of and
the reluctance to use maternity services are inextricably linked to
deep-rooted issues such as the state of the health system and the
place of women in society (Koblinsky et al., 2006). Because profes-
sional care at childbirth is often used as a broader proxy for the state
of a health system (Rohde et al., 2008), these trends raise concerns
beyond maternal health.
In response financial incentives have been increasingly advo-
cated as an effective means to change health-related behaviours
and improve health outcomes (NORAD, 2007). If households lack
the financial resources, heavily discount the future or lack infor-
mation on the benefits of health care to make optimal care seeking
choices, financial incentives can increase demand for health care.
∗ Corresponding author. Tel.: +44 020 7612 7887.
E-mail address: Timothy.Powell-Jackson@lshtm.ac.uk (T. Powell-Jackson).
Financial incentives are the key feature of various programmes that
have become popular in recent years, including conditional cash
transfers, vouchers and one-off cash payments.
Financial incentives provide an immediate reward to individ-
uals for behaviour that leads to health gains, and have been used
to target a range of health-related behaviours. Recent enthusiasm
for their use in low and middle income countries is supported
by evidence showing that payments aimed at initiating take up
of preventive health interventions can be effective (Lagarde et al.,
2007). However, there is also limited evidence of perverse effects.1
Financial incentives have been used in a positive sense, to encour-
age uptake of health technologies and attendance at health clinics
(Fiszbein et al., 2008). More controversially perhaps, they have been
used to encourage individuals to refrain from certain behaviours,
such as contracting sexually transmitted diseases (Jack, 2008).
In this paper we explore the effect of a national financial incen-
tive programme for maternal health in Nepal. Despite a recent
improvement in maternal mortality in the country, utilisation of
maternity services has remained unacceptably low (Pradhan et al.,
1
With a few notable exceptions, there is little evidence on undesirable effects
of financial incentives in health. Conditional cash transfers in Honduras may have
increased fertility because only pregnant women were eligible to benefit (Morris
et al., 2004a). It is argued that children in Brazil may have been kept malnourished
owing to a misperception that this would qualify the household for financial benefits
(Morris et al., 2004b).

0167-6296/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhealeco.2011.10.010
272 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284

Table 1 health services are available and government financial systems

Financial incentives offered by the SDIP.
Financial incentive
1. Cash payment to women
• 500 NRS ($7.8) in plains districts
• 1000 NRS ($15.6) in hill districts
• 1500 NRS ($23.4) in mountain
districts
2. Provider incentive
• 300 NRS ($4.7) for each delivery
attended
3. Free delivery care to women and
facility reimbursed
• 1000 NRS ($15.6) reimbursed to
health facility
Eligibility criteria
Woman delivered in a public health
facility and had no more than two
living children or an obstetric
complication (as diagnosed by the
health provider)
Doctor, nurse, midwife, health
assistant, auxiliary health worker or
maternal and child health worker
attended a delivery at the woman’s
home or in a public health facility
Woman comes from one of the 25 least
developed districts and meets the
eligibility criteria required for the
financial incentive
relatively strong. This paper also informs the debate on the
feasibility of implementing financial incentive programmes in
resource-poor settings. While the notion of paying individuals to
influence their behaviour is simple and intuitively appealing, our
findings suggest that such interventions can be complex to imple-
ment (Oxman and Fretheim, 2008).
The paper is structured as follows: Section 2 considers the main
theoretical mechanism underpinning the SDIP in formulating pre-
dictions as to its effect. Section 3 describes the methods, including
our definition of treatment, the empirical strategy and the data used
in the study. Section 4 presents the findings and Section 5 discusses
the main implications and limitations of the study.
2. Theoretical considerations

We start with a conceptualisation of the pathways through

1997; Government of Nepal, 2001, 2007). The Government of Nepal
thus turned to the use of financial incentives. Introduced nation-
wide in July 2005, the Safe Delivery Incentive Programme (SDIP)
provides: (i) a cash payment to women who give birth in a public
health facility; (ii) exemption from user fees for those residing in
the least developed one third of districts; and (iii) a financial incen-
tive to health workers. The incentive to health workers is given for
attendance at deliveries both in the health facility and at the home
of the woman giving birth (Government of Nepal, 2005). As shown
in Table 1, the amount of cash was designed to vary across the
three main geographical regions of Nepal to reflect differences in
the cost of accessing health services faced by households (Borghi
et al., 2006a). The development of the SDIP and its rapid adoption
was heavily influenced by a convergence of political interests and
effective dissemination of research findings supporting the notion
of financial incentives (Ensor et al., 2009). At the time, the coalition
government was headed by the United Marxist Leninist party who,
in their manifesto, had pledged support to advancing the status of
women.
We focus on estimating the effect of the SDIP on women’s use
of health care services at childbirth. The cost of maternity care
faced by households can be high, with the majority of expenditures
made outside of the health facility (Borghi et al., 2006b). By reduc-
ing these costs, the SDIP is expected to lead to improved health
seeking behaviour at childbirth. We estimate the magnitude of the
effect on use of formal care and then seek to understand whether
the benefits of the SDIP vary according to characteristics of the tar-
get population and the design of the programme. Variation in the
package of financial benefits across regions, for example, provides
an opportunity to explore whether the size of the incentive makes
any difference to the impact of the programme.
Our empirical strategy relies on an unusual measure of treat-
ment, namely the woman’s knowledge of the SDIP prior to
childbirth, and propensity score matching methods to estimate
the causal impact of the programme. Identification rests on the
strong assumption of conditional independence and, for this rea-
son we explore a number of approaches to assess the robustness
of the basic findings. Over our study period, implementation was
characterised by lengthy delays in the disbursement of funds from
the central level and hesitation on the part of the government to
promote the programme using mass media (Powell-Jackson et al.,
2009a). As we argue later, the extent of implementation must be
given consideration when interpreting the findings.
The paper contributes to the growing literature on demand-side
incentives in health. However, there is little rigorous evidence on
whether financial incentives work in low-income countries, partic-
ularly when implemented at scale. The available evidence comes
largely from middle-income Latin American countries, where
which the SDIP can be expected to improve outcomes, in an effort to
make explicit the assumptions that underpin the process (Weiss,
1998; White and Masset, 2007). This leads us to identify a num-
ber of steps that can be considered necessary if the programme
is to lead to a change in health seeking behaviour. It also allows
us to make the distinction between individual actions and govern-
ment involvement in the implementation process and provides the
basis with which we define our treatment group in the subsequent
analysis.
At the individual level, the key steps include: households with a
pregnant woman hear about the financial incentives offered by the
SDIP; households perceive the promise of financial incentives as
credible – that is, they expect to receive the benefits; women give
birth with professional care in a health facility; and after giving
birth women receive the demand-side incentives in a timely man-
ner. Women’s experience of the administration process is commu-
nicated to other families and this in turn, along with other factors,
affects their expectation of whether they will receive the demand-
side incentives in the future. The government’s role in the imple-
mentation process is to promote the SDIP if target households are
to know about the demand-side incentives on offer and to ensure
funds are available in health facilities if women are to be paid on
time. The latter requires a well functioning public financial man-
agement system since the funds must flow from the central trea-
sury to each district health office and then to each health facility.
In formulating predictions about the effect of the SDIP on health
care seeking, we emphasise the price mechanism as the primary
causal pathway through which the demand-side incentives affect
behaviour. A simple model of provider choice illustrates that an
individual is essentially faced with a trade-off between health and
non-health consumption (Gertler and Van der Gaag, 1990). The
SDIP’s demand-side incentive represents a subsidy on the price
of care which will induce an increase in the demand for publicly
provided maternity services. The increase in demand is the result
of two effects: a substitution effect and an income effect. The for-
mer occurs because of a change in the relative prices of alternative
providers, while the latter is the result of an increase in purchasing
power.
The extent to which the demand-side incentives increase the
use of formal care in the public sector (i.e. the price elasticity of
demand) is an empirical question and the main focus of this paper.
Since the SDIP operates only in the public sector, we anticipate a
substitution away from home care and non-state providers. It can
also be shown that poorer individuals have a lower price elasticity
of demand for health care than wealthier individuals, as long as
health is a normal good (Gertler et al., 1987). For the increase in
demand to translate into utilisation, maternity services must be
available and the quality of these services will determine whether
 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284
there are improvements in health outcomes. In other words, we
expect the effect of the demand-side incentives to be greater where
health services are available.
3. Methods
3.1. Defining treatment
Implicit in the archetypal evaluation problem with a binary
treatment is a clear definition of the treatment status of each indi-
vidual in the population of interest. The most common way of
defining treatment uses enrolment status, eligibility status or geo-
graphical placement of the programme. In this study, all three were
ruled out owing to the nature of the programme and the fact that
it was launched nationwide from the outset.
Instead, the study design is informed by the model of the
programme’s causal pathway. The implementation process points
towards knowledge about the SDIP prior to childbirth as a nec-
essary condition for the programme to affect health seeking
behaviour. A family’s decision of where to seek care may be influ-
enced by its ex-ante expectation of the price of care. It follows that
the SDIP’s demand-side incentives will only influence the health
seeking behaviour of those families who have knowledge of the
programme’s benefits before childbirth. This definition of treat-
ment implies that only the impact of the financial incentive and
free delivery care – and not the health provider incentive – can
be assessed. The impact estimates, therefore, do not reflect the
possible supply-side influence of the provider incentive on utilisa-
tion, although we do recognise that the impact of the demand-side
incentives may vary according to supply-side factors.
Household interviews were carefully conducted to ensure the
knowledge of the women and the family decision-maker was cap-
tured to reflect the fact that few married women in Nepal make
decisions regarding their own health care (Government of Nepal,
2007).2 There were concerns that the SDIP might be confused with
other health programmes and the use of ex post information on
knowledge of the programme prior to childbirth might be suscep-
tible to recall bias. Thorough pre-testing of the survey tool and
discussions with respondents after each interview provided reas-
surance that the validity of the measure was not compromised
by these two concerns. Specifically, women found the SDIP’s offer
of cash a highly distinctive feature. This meant that there was
little risk of respondents confusing the SDIP with other health
programmes since no other government programme offers cash
to women. Furthermore, childbirth is a highly memorable event
(whether it is positive or negative), allowing women to relate the
timing of when they found out about the SDIP to the date of giv-
ing birth.3 With treatment measured in this way, the survey tool
aimed to collect information on factors that would be expected to
influence jointly exposure to information about the SDIP and health
seeking behaviour at childbirth.
3.2. Empirical strategy
We identify the impact of the SDIP using propensity score
matching methods. The idea behind the approach is to select a com-
2
We refer to the woman’s knowledge of the SDIP throughout the paper, but this
should always be interpreted as the family’s knowledge of the SDIP.
3
The most common misreporting is expected to be women stating that they knew
about the SDIP during pregnancy when in fact they found out subsequently. This
would downward bias impact estimates since these women, incorrectly classified
as treated individuals, were not exposed to information about the SDIP and are thus
less likely to have delivered in a health facility.
273
parison group of non-participants that is as similar as possible to
the treatment group in its observed characteristics. Individuals in
the comparison group are selected on the basis of their propen-
sity score, given by P(Z) = Pr(D = 1|Z) and (0 < P(Z) < 1), where D is a
treatment indicator and Z is a set of control variables unaffected
by programme participation. Appropriate matching variables are
those that jointly affect treatment status and the outcome. The
propensity score gives the probability that an individual partici-
pates in the programme given the set of observed characteristics
that jointly influence treatment status and outcomes. It is estimated
by means of a probit model.
Two assumptions are required for the identification of the aver-
age treatment effect on the treated (ATT). First, the conditional
independence assumption, or unconfoundedness, states that out-
comes are independent of participation given the observables (i.e.
Y0 , Y1 ⊥ D|Z). It can be shown that the conditional independence
assumption continues to hold conditional on P(Z), such that out-
comes are independent of participation given the propensity score
(i.e. Y0 , Y1 ⊥ D|P(Z)) (Rosenbaum and Rubin, 1983, 1984). Second,
the common support assumption ensures that there are treated and
untreated individuals with the same characteristics (P = D = 1|Z < 1).
When these conditions hold, the average treatment effect on
the treated is identified non-parametrically by the mean condi-
tional difference in the outcome over the common support, suitably
weighted by the distribution of Z in the treatment group. There
are a number of ways to construct the matched outcome that
vary around how the set of neighbours, C0 (Zi ), is defined and how
the weights, Wij , are chosen. We try three matching estimators
and evaluate how well they balance the covariates.4 With caliper
matching, each treated observation is matched to the 10 neigh-
bours nearest in terms of their propensity score (Cochran and
Rubin, 1973). Untreated observations must fall within a maximum
distance (caliper) of 0.01 in order to be matched. In kernel match-
ing, the outcome of a treated individual is matched to the weighted
outcomes of all untreated units, where the weight is in propor-
tion to the closeness between the propensity score of the treated
and untreated individuals (Heckman et al., 1998). The bandwidth
is set at 0.03. Finally, Mahalanobis-metric matching combines the
matching variables into a distance measure and then matches
based on the resulting scalar. The propensity score is included in
the set of matching variables.
Although we use a rich set of data from a survey designed specif-
ically to evaluate the SDIP and capture information on some of
these ‘unobservables,’ conditional independence remains a strong
assumption and the basic results are unlikely on their own to pro-
vide fully convincing evidence of a causal effect. We pursue three
strategies to mitigate concerns about bias due to potential correla-
tion between exposure to the SDIP and unobserved factors affecting
health seeking behaviour. First, we analyse the impact of the SDIP
by the type of provider to explore whether there is any evidence
of a substitution effect. If the ATT estimates are subject to omitted
variable bias, this bias can be expected to work in the same direc-
tion for state and non-state health providers, given that they are
4
The percent of treated observations lost due to common support and various
tests of matching quality provide the basis with which to evaluate the matching
procedures. Smith and Todd (2005) suggest testing for differences in the covariates
between the treated and the non-treated group. For each variable, the standardised
percentage bias – the difference of the sample means in the treated and non-treated
groups as a percentage of the square root of the average of the sample variances in
each group (Rosenbaum and Rubin, 1985) – is calculated before and after matching.
A two-sample t-test can then be used to check if there are significant differences
between the means before and after matching. In addition, a likelihood-ratio test
of joint significance of covariates before and after matching can provide tests of
matching quality.
274 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284
the closest substitutes. Impact estimates that are upward biased
are likely to hide all evidence of a substitution effect between the
different types of provider.
A second strategy is akin to providing evidence on a
dose–response relationship. It builds on the insight that knowl-
edge of the SDIP is necessary but insufficient if the programme is
to have an impact. There must also be a strong expectation that the
woman will receive the cash incentive after childbirth. We rede-
fine treatment to create two new treatment groups. TCT is made
up of women who knew about the SDIP prior to childbirth and
expected to receive the cash. TPT consists of women who knew
about the SDIP but did not expect to receive the cash incentive.
Expectations of receiving the cash are measured by whether the
woman knew of anyone else who had received the cash incen-
tive of the SDIP.5 Propensity score matching is applied to each of
these treatment groups in separate analyses using the same com-
parison group as previously. Given that we expect the treatment
effect in the analysis with TPT as the treatment group to be zero
coupled with the fact that women in the comparison group and
treatment group have very different characteristics (and different
biases), any omitted variable bias should be apparent through a dif-
ference in the mean utilisation of maternity services between the
two groups (Imbens and Wooldridge, 2009). In this sense, a find-
ing of no treatment effect in the analysis with TPT as the treatment
group, and a large treatment effect in the analysis with TCT as the
treatment group provides evidence in support of the assumption
of conditional independence.
Third, we control for unobservables using an instrumental vari-
able strategy to explore whether omitted variable bias plagues the
propensity score estimates. A promising candidate to instrument
knowledge of the SDIP during pregnancy is the frequency with
which women in the sample listen to the radio. We know from
the survey that some women heard about the SDIP from the radio
and indeed district health offices reported separately in qualitative
interviews that they used the radio to promote the programme.
The radio was not used, however, to promote institutional delivery
care more generally.6 Thus, to the best of our knowledge, the only
pathway through which the radio influences a woman’s place of
delivery is through its promotion of the SDIP. The two-stage least
squares estimate of impact should be interpreted as a local average
treatment effect, given that the assumption of homogeneous treat-
ment effects is particularly unrealistic in this instance (Imbens and
Angrist, 1994). The treatment effect, in other words, applies only
to the sample of women who found out about the SDIP through the
radio.
A final analysis explores heterogeneity in the impact of the SDIP
(Djebbari and Smith, 2008). The effect of the programme is antici-
pated to vary along a number of dimensions, including household
wealth, the level of benefits offered in a particular district and
the availability or quality of care in the public sector. We analyse
heterogeneity in impacts across our observables using subgroup
analysis, with a view to disentangling which factors are associated
with variation in the ATT estimates. Variation in impact by wealth,
for example, may be due to a correlation between a household’s
5
The rationale for this measure is that a person’s perception of the administra-
tion of the programme is particularly sensitive to the experiences of others in the
community. Clearly, this indicator is a crude measure of expectations given that it is
binary and fails to capture the full range of values from a probability. Nevertheless,
it seems reasonable that expectations of receiving the cash incentive will be higher
among those who knew of someone who had received the money than those who
had never heard of anyone.
6
The evaluation of the SDIP used research teams drawn from each of the study
districts. These individuals had a deep knowledge of the local area and were able to
confirm that FM radio was used solely to promote the SDIP.
wealth and the level of benefits provided in its district when it is in
fact the latter which drives the relationship (Wagstaff et al., 2009).
In a linear probability model, we regress the utilisation outcome
on our treatment variable, the covariates and a set of interactions
between treatment and the covariates. The regression is weighted
by the kernel weight (Leuven and Sianesi, 2003).7
We include in the analysis of heterogeneous impacts a measure
of the availability of maternity care for each type of government
health facility.8 We consider this measure a proxy for quality of
care. In a survey of health providers conducted in parallel with the
household survey, we asked each health facility whether it had pro-
vided a number of services that are considered critical components
of maternal health care (UNICEF, 1997). Specifically, we asked if
the health facility had (in the last three months): (i) administered
parenteral antibiotics; (ii) administered parenteral oxytocics; (iii)
administered anticonvulsants; (iv) performed manual removal of
placenta; (v) performed manual removal of retained products; (vi)
performed assisted vaginal delivery; (vii) given a blood transfu-
sion; (viii) performed a caesarean section; (ix) referred a woman
by ambulance; (x) provided 24 h delivery care services. The mea-
sure we use is the number of procedures that the health facility was
able to carry out (0–10). The variables are measured at the district
level, taking the mean value if there is more than one facility of that
type in the district.
3.3. Data
The data used in the analysis come from a survey of women
that was conducted two and a half years after the start of the SDIP.
Women who had given birth in the previous three years were inter-
viewed across six districts of Nepal. Two districts were randomly
selected from each of three ecological regions of Nepal to ensure we
had variation in the level of benefits offered by the SDIP across the
districts. The sample was selected in two stages: in the first stage
180 primary sampling units (villages in rural areas or urban wards)
were selected using systematic sampling with probability propor-
tional to size, and in the second stage an average of 30 women in
each village were chosen randomly.
Each observation is a delivery and, given the three year recall
period, it is possible for a woman to have had more than one deliv-
ery. The recall period extended five months before the start of the
SDIP. We dropped two observations for lack of data. The dataset
contains complete data on 5903 deliveries, of which 420 took place
before the start of the SDIP. In the sample, women were unaware of
the SDIP prior to childbirth in three-quarters of cases. These women
make up our comparison group.
The impact of the SDIP is assessed on a set of binary utili-
sation outcomes that refer to the place of delivery, the type of
attendant present at the delivery and the type of procedure,9 if
carried out during the delivery. Utilisation of professional delivery
care services in the comparison group is low (Table 2). Deliver-
ies in a health facility account for 16 percent of all deliveries and
the public sector is the dominant provider. Just over 10 percent
of deliveries in the comparison group take place in a govern-
ment health facility, while non-governmental (not-for-profit) and
7
As noted by Wagstaff et al. (2009), the standard errors from the regression are
smaller than those generated by psmatch2, which means in our analysis of hetero-
geneous effects we report t-statistics that are somewhat higher than warranted.
8
Through its supply-side incentives, the SDIP may have improved the availabil-
ity of maternity services and thereby increased use of maternal health care. This
possibility, however, should not threaten the internal validity of our main findings
which isolate the demand-side effect of the programme.
9
An assisted delivery refers to the use of forceps or a ventouse that are attached
to the baby’s head.
 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284 275

Table 2
Descriptive statistics, by treatment group.

Variable Treated Comparison

Mean Std deviation Mean Std deviation

Outcomes
Health facility 0.263 0.440 0.155 0.362
Government health facility 0.210 0.407 0.106 0.307
NGO hospital 0.024 0.154 0.031 0.175
Private health facility 0.029 0.168 0.018 0.134
Doctor, nurse or midwife in attendance 0.293 0.455 0.182 0.386
Any health worker in attendance 0.351 0.478 0.225 0.418
Caesarean section 0.047 0.212 0.025 0.155
Caesarean section or assisted 0.099 0.299 0.048 0.214

Covariates
Age of woman 25.379 5.285 26.622 5.990
Log of wealth 3.062 0.658 2.911 0.676
No education (reference)
Primary education 0.135 0.342 0.119 0.323
Secondary education 0.271 0.445 0.195 0.396
Higher education 0.189 0.391 0.080 0.272
No work (reference)
Agriculture work 0.684 0.465 0.701 0.458
Salaried work 0.030 0.169 0.012 0.109
Small business 0.066 0.248 0.046 0.209
Waged work 0.063 0.243 0.084 0.278
Other work 0.003 0.058 0.002 0.045
Brahmin and Chhetri (reference)
Terai and Madeshi 0.063 0.243 0.079 0.270
Dalit 0.118 0.323 0.141 0.348
Newar 0.043 0.203 0.016 0.125
Janajati 0.173 0.379 0.204 0.403
Muslim 0.005 0.068 0.028 0.165
Other castes 0.075 0.263 0.129 0.335
Walk to facility < 1 h (reference)
Walk to facility 1 h < 4 h 0.613 0.487 0.572 0.495
Walk to facility 4 h < 1 day 0.115 0.319 0.184 0.388
Walk to facility > 1 day 0.031 0.173 0.106 0.307
Urban dwelling 0.120 0.325 0.084 0.277
Previous delivery during SDIP 0.091 0.287 0.060 0.238
Active FCHV 0.147 0.128 0.076 0.098
Women’s groups 0.337 0.473 0.274 0.446
Morang (reference)
Sankhuwasabha 0.238 0.426 0.140 0.347
Myagdi 0.163 0.370 0.142 0.350
Rupandehi 0.109 0.312 0.160 0.367
Jumla 0.217 0.412 0.185 0.389
Achham 0.129 0.335 0.189 0.391
Year 1.947 0.752 1.617 0.872
Observations 1489 4416

Note: FCHV = female community health volunteer.

private (for-profit) health facilities account for 3 percent and 2 per-
cent of all deliveries respectively. Almost a fifth of women in the
comparison group give birth with a doctor, nurse or midwife in
attendance, the indicator that corresponds with the standard inter-
national definition of a skilled birth attendant. The rate of caesarean
section in the study area is low, with only 3 percent of women
having surgery at childbirth. While the survey was not designed
to be representative of the entire country, the comparison group
estimates correspond very closely with national estimates, which
mostly covered the period before the SDIP (Government of Nepal,
2007).
The decision of what covariates to include is informed by the
various sources of information from which women found out about
the SDIP. Table 2 provides summary statistics of the covariates
used in the propensity score matching. Most, but not all, of the
covariates are self-explanatory. A wealth index is constructed from
information on the ownership of assets using principal compo-
nents analysis and re-scaled to give values between 1 and 100
(Filmer and Pritchett, 1998). We take the log of the re-scaled index.
We include two village level variables. The level of activity of
the female community health volunteer is an index that ranges
between possible values of 0–1. It provides a measure of how
active the female community health volunteer in each village is
in disseminating information about the SDIP to the community. It
is calculated as the proportion of women in a village who found
out about the SDIP through the female community health volun-
teer. Each observation takes the mean value for the village. The
women’s group variable indicates whether the village in which
the woman lives has women’s group meetings. The year vari-
able takes the value of the fiscal year in which the delivery took
place with possible values ranging from 1 to 4, and 2 represent-
ing the first year of the SDIP. It captures unobservables whose
variation over time influence both programme participation and
outcomes.
Based on these covariates, the two groups of women appear to
be different. The comparison group is older, less wealthy, less edu-
cated, more likely to be from marginalised castes, and lives further
from its nearest health facility than the treatment group.
276 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284

Table 3
Covariate balancing indicators before and after matching.

Matching estimator N1 N0 Probit pseudo R2 Probit pseudo R2 p > 2 Mean bias (%) Mean bias (%) Lost to common support (%)
Before Before Before After After Before After After

Kernel matching 1489 4416 0.177 0.002 1.000 17.18 1.74 0.13
Malahanobis matching 1489 4416 0.177 0.021 0.000 17.18 3.46 0.00
Nearest neighbour matching 1489 4416 0.177 0.002 1.000 17.18 1.88 0.81

Note: The pseudo R2 is from the probit estimation of the conditional treatment probability and gives a measure of how well the regressors explain variation treatment. The
p value of the likelihood ratio test after matching tests the hypothesis that the regressors are jointly significant. The mean bias is the mean absolute standardised bias, as
defined in Section 3.2. Lost to common support gives the percentage of treated observations that fall outside the boundaries.

3.4. Propensity score estimation and balancing
The results of the probit model are shown in
Table A1 in the Appendix. While the probit model is primar-
ily used as a statistical tool to estimate the propensity score, it can
also shed light on inequalities in implementation. The pseudo R2 of
the probit model is 0.177 and the coefficients indicate that a hand-
ful of covariates are important predicators of treatment status.
Younger, more educated, and wealthier women are more likely to
have knowledge of the SDIP during pregnancy. Women belonging
to Janajati or Muslim castes are less likely than the reference group
(Brahmin/Chhetri) to have knowledge of the programme, while
those living closer to a health facility are more likely to have heard
of the programme, as are women living in villages with active
female community health volunteers. If a woman gave birth to a
previous baby after the start of the SDIP, she has a higher chance
of knowing about the programme. Women from the districts of
Sankhuwasabha, Myagdi, Rupandehi and Achham are more likely
than those in Morang to know about the SDIP while they were
pregnant. Lastly, as the programme has matured, exposure to
information on the programme has increased, as indicated by the
effect of the year variable.
The purpose of the matching exercise is to balance the covariates
such that the bias on observables is reduced as much as possible.
Overall measures of imbalance in the covariates before and after
matching are reported in Table 3. The pseudo R2 from the probit
model estimated on the trimmed sample using the weights gener-
ated from all three matching exercises is appreciably lower than the
value from the model on the original unweighted sample. Kernel
matching and nearest neighbour matching perform best, reducing
the pseudo R2 value from 0.177 to 0.002 respectively.
The hypothesis of the joint insignificance of all the regressors
before and after matching cannot be rejected in the case of the
kernel matching and the nearest neighbour matching, as indicated
by the p-values of the likelihood-ratio test. Before matching, the
average standardised bias is 17.2 percent. All three matching meth-
ods reduce the bias dramatically, although some more than others.
without the need to discard a large number of observations from
the sample due to a lack of common support. Since the region of
common support includes almost the entire sample, the estimated
treatment effect barely has to be redefined, essentially allowing
the ATT to be recovered. While there is little to choose between the
kernel and the nearest neighbour matching procedures, the former
is preferred on the basis that it loses the least observations to com-
mon support while achieving the greatest reduction in bias. The
next section reports estimates of programme impact using kernel
matching.
4. Results
4.1. Effectiveness of implementation
To provide some context behind the impact results, we first
examine how well the programme was implemented. Two mea-
sures are particularly revealing. The first concerns awareness of
the SDIP among the target population. Just under a quarter (24.3
percent) of women in our sample had knowledge of the SDIP prior
to childbirth. This estimate implies that three-quarters of women
were not reached by the programme and their health seeking
behaviour could not plausibly have been influenced by the offer of
the financial incentive. In Fig. 2, the concentration curve of knowl-
edge of the SDIP shows evidence of some inequality, particularly in
the poorest wealth quintile.
The second measure of implementation concerns receipt of the
financial incentive. Only 26.5 percent of women who gave birth in a
government health facility were given the financial incentive or an
equivalent deduction on their bill. There were clearly considerable
problems in paying women. The vast majority of women who were
meant to receive the financial incentive were never paid. Inequal-
ity in the receipt of the financial incentive across the entire sample
Untreated (never heard of SDIP)
4

Kernel matching reduces the average bias the most to 1.7 percent,
while caliper matching reduces the average bias to 1.9 percent.
3

Mahalanobis matching performs less well, reducing the average

2

bias to 3.5 percent. Further evidence on the quality of matching is

1

provided in Table A2 in the Appendix.

The percentage of observations lost to the common support
Density
0

Treated (heard of SDIP)

restrictions is negligible indicating that none of the matching
4

methods appear to pose a problem in this regard. The kernel, Mala-

hanobis and caliper matching procedures lose two (0.13%), zero and
3

twelve (0.81%) treated observations respectively. Fig. 1 shows the

2

histogram of the propensity scores before matching for those in the

1

treatment group and those in the comparison group. The region of

common support is substantial despite there being a right-skewed
0

0 .5 1
distribution in the case of the untreated deliveries. Pr(TREAT)
The common support and balancing test results suggest that Graphs by psmatch2: Treatment assignment

bias associated with differences in the observables between the

treated and untreated groups can be almost completely eliminated Fig. 1. Histogram of propensity scores for the treated group and comparison group.
 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284 277

1 Actions at the district level appeared to be have been influenced

by the pressure to meet local needs, as well individual perceptions
Cumulave share of women reached by SDIP

0.9 Received the cash incenve

Knew about SDIP and acceptance of the programme among district implementers.
0.8

0.7 Line of equality

4.2. Main impact results
0.6

0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3
Cumulave share of deliveries (poorest first)
Fig. 2. Concentration curves for knowledge of the SDIP prior to childbirth and
receipt of the financial incentive.
of women gives an indication of the benefit incidence. As the con-
centration curve in Fig. 2 shows, the financial incentive is captured
disproportionately by relatively wealthier households. While this
largely reflects inequality in the use of government health services,
it does serve to show that without any targeting such an interven-
tion effectively subsidises the care of wealthier households who
use services.
Findings from a process evaluation, conducted in parallel with
this study shed light on why implementation in the first few years
of the SDIP was so incomplete (Powell-Jackson et al., 2009a). The
key implementation challenges included bureaucratic delays in the
disbursement of funds and difficulties in communicating the policy
to the public and health providers. There were lengthy delays in the
transfer of funds from the central government to the districts. Dis-
tricts on average received funds earmarked for the SDIP 283 days
late in the first fiscal year, and 147 days late in the second fiscal year,
with much of the delay caused by the late transfer of funds from the
international donor to the central government. Because the cen-
tral government was then wary of raising expectations among the
public without having the funds to pay beneficiaries, programme
managers chose not to publicise the SDIP with a national media
campaign.
These two problems at the central level provide a convincing
explanation for the broad findings presented above. It should be
noted, however, that some districts were able to cope remark-
ably well with these challenges, providing an explanation for why
uptake of the programme varied considerably between districts.
Table 4 shows estimates of the ATT for each utilisation out-
come with the corresponding t-statistic, the percentage change
relative to the mean of the untreated and the 95% confidence
interval around the relative change. In column 1, estimates before
matching indicate that there are substantial differences in utilisa-
tion between the treated and untreated groups. However, all this
serves to show is that there is likely to be selection bias on the
observables.
0.4 0.5 0.6 0.7 0.8 0.9 1
The matched estimates indicate that the SDIP, for those it
reached, had a statistically significant effect on the probability of
women delivering in a health facility. Institutional deliveries for
those treated increased by 4.0 percentage points (relative change
18%; CI: 5%, 31%) as a result of the SDIP. Among those reached by
the SDIP, impact estimates show that the programme increased
both deliveries with a skilled birth attendant and deliveries with
any professional health worker by a statistically significant 4.2 per-
centage points (relative change 17%; CI: 4%, 29%) and 5.2 percentage
points (relative change 17%; CI: 6%, 28%), respectively.
The programme also had a positive effect of 1.2 percentage
points (relative change 36%; CI: −3%, 74%) on the caesarean section
rate (significant at the 10 percent level), and a positive impact of 1.9
percentage points (relative change 24%; CI: 0%, 48%) on caesarean
sections and assisted deliveries combined. The results are robust
to the matching procedure used (see Table A3 in the Appendix).
There is consistency in the statistical significance and magnitude
of impact estimates across the different outcomes. Programme
impacts tend to be greatest with the nearest neighbour matching.
The findings provide encouraging signs that the SDIP increased
use of formal maternal health care. The positive effect of the SDIP
on the caesarean section rate is more difficult to interpret because
there is no way of knowing whether the additional procedures were
medically required. But, given how low the caesarean section rate
is in Nepal and the fact that there are no obvious incentives to the
provider for performing more (unnecessary) caesarean sections,
this result is likely to reflect an improvement in welfare.
4.3. Impact by type of provider
Any substitution effect from the reduction in the price of govern-
ment maternity services should become apparent when we analyse

Table 4
Impact of the SDIP on health seeking behaviour.

Before matching After matching

Mean difference t-Stat ATT t-Stat ATT as % of mean of untreated 95% confidence interval

Place of delivery
Health facility 0.107 9.33 0.040*** 2.70 17.8 (5.1; 31.1)

Type of attendant
Doctor, nurse or midwife (SBA) 0.110 9.11 0.042*** 2.72 16.6 (4.1; 29.1)
Any professional health worker 0.126 9.70 0.052*** 3.17 17.2 (6.1; 28.1)

Procedure at delivery
Caesarean section 0.022 4.35 0.012* 1.83 35.5 (−3.1; 74.1)
Caesarean section or assisted 0.051 7.15 0.019** 2.02 23.9 (0.1; 48.1)

Note: Sample includes 5901 deliveries. t-Statistics are based on standard errors provided by psmatch2. Kernel matching with a bandwidth of 0.03 is used to compute impact
estimates. A treated observation whose propensity score is greater than the largest of the propensity scores of the untreated is dropped. 95% confidence intervals are around
the relative percentage impact. ATT = average treatment effect on the treated. SBA = skilled birth attendant.
*
Impact parameters are significant at 10%.
**
Impact parameters are significant at 5%.
***
Impact parameters are significant at 1%.
278 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284
Table 5
Impact of the SDIP on health seeking behaviour by type of health facility.
ATT t-Stat
***
Health facility 0.040 2.70
Government health facility 0.043*** 3.28
NGO hospital −0.011* −1.86
Private health facility 0.008 1.40
Note: Sample includes 5901 deliveries. t-Statistics are based on standard errors provided by psmatch2. Kernel matching with a bandwidth of 0.03 is used to compute impact
estimates. A treated observation whose propensity score is greater than the largest of the propensity scores of the untreated is dropped. 95% confidence intervals are around
the relative percentage impact. ATT = average treatment effect on the treated. NGO = non-governmental organisation.
*
Impact parameters are significant at 10%.
**
Impact parameters are significant at 5%.
***
Impact parameters are significant at 1%.
the impact of the SDIP on utilisation by type of provider. Table 5
presents the ATT estimates for various types of health facility. The
impact of 4.3 percentage points (relative change 26%; CI: 10%, 42%)
in the government sector is offset by a negative effect of 1.1 per-
centage points (relative change −31%; CI: −65%, 2%) on utilisation
of NGO hospitals. These results suggest that the SDIP encouraged
women to switch from NGO hospitals to government health facil-
ities. Reassuringly, this substitution effect did not account for the
entire increase in the utilisation of government providers and, as
noted in Table 4, there remains a statistically significant net positive
impact on institutional delivery care.
Intuitively, this finding makes sense. Government and NGO
health providers are close substitutes and it seems plausible that a
subsidy targeting only the public sector encourages some women
to switch from the NGO sector. The private for-profit sector, on the
other hand, caters to only the wealthiest who are less likely to be
incentivised by the offer of the SDIP’s cash.
This finding also serves to strengthen the robustness of the basic
results. Since health facilities in the public and NGO sectors are
close substitutes, any bias on unobservables can be expected to
operate in the same direction for both types of provider. Thus, if
omitted variable bias was driving the positive impact on utilisation
of government health facilities, it seems unlikely that there would
be any evidence of a negative effect on deliveries in NGO hospitals.
4.4. Credibility of the SDIP’s promise of cash
In the following analysis, we assume that it is not sufficient to
have knowledge of the SDIP for the programme to influence health
seeking behaviour. There must also be a strong expectation that the
cash incentive will be given to the woman after childbirth. Table 6
reports the findings from the two analyses in which the treatment
group is defined differently.10 In column 1, the ATT estimates sug-
gest that the SDIP had a large and significant impact on women
who knew about the SDIP and had heard of someone receiving the
financial incentive. These women increased utilisation of govern-
ment maternity services by 9.6 percentage points (relative change
53%; CI: 31%, 74%) because of the SDIP.
In column 5, the results suggest that knowing about the SDIP
while having little expectation of receiving the cash incentive has
no effect on utilisation of maternity services. Knowledge on its
own does not appear sufficient to change behaviour. The impact
estimates on the key outcomes are small in magnitude and highly
insignificant. A woman in this treatment group is 0.3 percentage
points (relative change −2%; CI: −21%, 17%) less likely to deliver
in a government health facility as a result of the SDIP and 0.2 per-
centage points (relative change −0.9%; CI: −15%, 13%) less likely to
10
In terms of the balancing of the covariates, we hasten to add that the kernel
matching estimator performs well in both matching exercises.
ATT as % of mean of untreated 95% confidence interval
17.8 (5.1; 31.1)
25.8 (10.1; 42.1)
−31.4 (−65.1; 2.1)
36.3 (−16.1; 88.1)
deliver with skilled birth attendance. In columns 4 and 8 we can
see the 95% confidence intervals of the treatment effects from the
two sets of results do not overlap for these outcomes.
The results from this analysis are in line with our expectation
and can be interpreted as providing evidence in support of the
robustness of the basic findings. The finding of no treatment effect
and the fact that the pseudo treatment group (TPT ) and the com-
parison group are very different from each other (and therefore
likely to have different biases) suggest that the conditional inde-
pendence assumption is plausible. If conditional independence did
not hold, the ATT estimates from the analysis of the pseudo treat-
ment group (TPT ) would most likely be biased upwards, showing a
spurious positive impact on utilisation.
The magnitude of the effect of the SDIP depends not only
the quality of implementation, as suggested by these results, but
also on prior expectations of receiving the financial incentive.11
If individuals at the outset have an overly optimistic view of the
government and over time are made to revise their expectation of
receiving the financial incentive down, there may be longer term
implications for the success of the SDIP and indeed other govern-
ment policies. We can examine this possibility by estimating the
effect of the SDIP over time, interacting six-monthly splines with
the treatment indicator. Although the coefficients on these inter-
actions are not significant, the pattern suggests that the effect of
the SDIP was greatest in the first six months of the programme and
diminished subsequently (result not shown).12
4.5. Instrumental variable estimates
To further address concerns over endogeneity, we instrument
knowledge of the SDIP prior to childbirth using the frequency of
radio listenership. Table 7 reports the impact estimate on util-
isation of government maternity services and for the sake of
convenience produces ordinary least squares (columns 1) and
propensity score matching (columns 2) results.
Column 3 confirms that the instruments are correlated with
knowledge of the SDIP prior to childbirth (F-statistic is 8.76).
The Sargan test of over-identification passes comfortably (Sargan,
1958). The impact estimates using two-stage least squares are,
if anything, greater in absolute size than those obtained using
propensity score matching and ordinary least squares. These results
suggest the propensity score matching estimates may not be sub-
ject to upward bias owing to omitted variables. The two-stage
least squares estimates of the effect of the SDIP are not, however,
11
We thank an anonymous for pointing this out to us.
12
This finding is also consistent with the results presented in Powell-Jackson et al.
(2009b), which uses longitudinal data from a community surveillance system in one
district to show that use of maternity care “jumped” at the start of the SDIP despite
the fact that disbursement of the financial incentive in the first few months of the
programme was particularly patchy.
 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284 279

Table 6
Credibility of the SDIP’s promise of the cash incentive.

TCT as treatment group TPT as treatment group

ATT t-Stat ATT as % of mean 95% confidence ATT t-Stat ATT as % of mean of 95% confidence
of untreated interval untreated interval
(1) (2) (3) (4) (5) (6) (7) (8)

Place of delivery
Health facility 0.094*** 4.36 38.3 (20.7; 55.8) −0.008 −0.47 −3.8 (−20.1; 12.1)
Government health facility 0.096*** 4.87 52.6 (31.0; 74.2) −0.003 −0.21 −2.0 (−21.1; 17.1)
NGO hospital −0.015* −1.85 −38.5 (−80.2; 3.1) −0.008 −1.19 −24.6 (−66.1; 17.1)
Private health facility 0.012 1.48 52.6 (−18.5; 123.7) 0.003 0.55 16.8 (−44.1; 78.1)

Type of attendant
Doctor, nurse or midwife (SBA) 0.090*** 4.06 33.0 (16.7; 49.2) −0.002 −0.13 −0.9 (−15.1; 13.1)
Any professional health worker 0.103*** 4.45 31.7 (17.5; 46.0) 0.006 0.33 2.2 (−11.1; 15.1)

Procedure at delivery
Caesarean section 0.023** 2.16 57.3 (4.2; 110.3) 0.002 0.27 6.1 (−39.1; 51.1)
Caesarean section or assisted 0.045*** 3.04 49.9 (17.1; 82.8) −0.001 −0.10 −1.5 (−32.1; 29.1)

Note: Sample includes 5093 deliveries in the comparison of treated and control observations. Sample includes 5223 deliveries in the comparison of pseudo control and
control observations. t-Statistics are based on standard errors provided by psmatch2. Kernel matching with a bandwidth of 0.03 is used to compute impact estimates. A
treated observation whose propensity score is greater than the largest of the propensity scores of the untreated is dropped. 95% confidence intervals are around the relative
percentage impact. ATT = average treatment effect on the treated. SBA = skilled birth attendant. NGO = non-governmental organisation.
*
Impact parameters are significant at 10%.
**
Impact parameters are significant at 5%.
***
Impact parameters are significant at 1%.

Table 7 that the coefficients on all the interaction terms are equal to zero,
Dealing with endogeneity concerns.
implying that the treatment effect is heterogeneous. The effect of
Method OLS PSM 2SLS wealth on the ATT estimate is negative but the coefficient is not
significant at any level. Inclusion of wealth quintile dummy vari-
(1) (2) (3)
ables does not alter this result. The impact of the programme does,
*** ***
Knowledge of SDIP 0.038 0.043 0.046
however, vary significantly by region, as shown by the rejection of
(0.011) (0.013) (0.153)
Controls Yes Yes Yes the null hypothesis that the region dummy variables are equal to
F test on instruments in first stage 8.76 zero.
Over-identification test p value 0.64 In the second model, the region dummy variables are replaced
Adjusted R2 0.21 – 0.22 by the size of the SDIP’s financial benefit as a share of the total
Observations 5903 5903 5903
cost of delivery to the household. The coefficient on the interac-
Note: Standard errors are reported in parentheses (robust standard errors in tion between this variable and treatment is positive and significant,
columns 1and 3). Controls include age, wealth (fourth degree polynomial), educa-
suggesting that the higher the subsidy relative to the cost of care,
tion, occupation of household head, caste, distance to nearest health facility, urban
residence, previous delivery during SDIP, active FCHV index, availability of women’s the greater the effect of the SDIP on utilisation. In the third model,
groups, and district fixed effects. The three instruments used in column 3 are: (i) we introduce the proxy measure of quality of care for each type
listens to radio daily; (ii) listens to radio at least once a week; (iii) listens to radio of government health provider. The effect of the size of the SDIP’s
less than once a week. The over-identification test is due to Sargan (1958). financial package is now even greater. We find that the quality of
*
Impact parameters are significant at 10%.
**
Impact parameters are significant at 5%.
care at hospitals and primary health care centres is positively asso-
***
Impact parameters are significant at 1%. ciated with the impact of the SDIP on utilisation of government
delivery care services. The quality of care in hospitals modifies the
effect the most, which seems plausible given that the majority of
close to being statistically significant. The loss of efficiency and women who use the public sector deliver in hospitals. While these
subsequently large standard errors are probably the result of two results are far from conclusive, they do provide some support to
problems. First, there is multicollinearity between the instruments the common sense notion that demand-side incentives are more
and several of the covariates, particularly education, which means effective when the supply of health services are reliably in place.
the variation left over in the error term in the first stage is vastly
reduced (Angrist and Pischke, 2009). Second, the fact that the 5. Discussion
dependent variable is binary increases the risk of a type II error
through lack of statistical power. This paper finds that the SDIP had a positive impact on the util-
isation of maternity services. Women in the treated group were
4.6. Heterogeneity in impacts 4.3 percentage points (26 percent) more likely to deliver in a pub-
lic health facility, 4.2 percentage points (17 percent) more likely
We analyse variation in the impact of the SDIP on utilisation to deliver with a skilled birth attendant and 1.2 percentage points
of government delivery care services. For various sets of variables, (36 percent) more likely to have a caesarean section. There was
we test the null hypothesis that the coefficients of the interaction evidence, albeit weaker, that the SDIP encouraged women to sub-
terms are equal to zero. Table 8 shows the results from three speci- stitute from NGO hospitals to give birth in government health
fications of a linear probability model. In the first model, we include providers. The SDIP, in other words, may have led to some crowding
the full set of covariates used in the estimation of the basic results, out of the private not-for-profit sector.
but rather than having district dummy variables we instead include While there is considerable enthusiasm for the use of demand-
region dummy variables to capture, inter alia, differences in the side incentives in resource-poor settings, it has been noted that
package of financial benefits. We are able to reject the hypothesis they may be inappropriate if health infrastructure is inadequate
280 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284

Table 8
Variation in the impact of the SDIP on utilisation of government maternity services.

Variable (1) (2) (3)

Coefficient Std error Coefficient Std error Coefficient Std error

Treatment 0.131 0.172 −0.087 0.134 −1.204** 0.468

Treatment × age of woman 0.002 0.002 0.003 0.002 0.002 0.002
Treatment × log of wealth −0.018 0.036 0.021 0.028 −0.040 0.036
Treatment × primary education 0.025 0.033 0.023 0.033 0.015 0.035
Treatment × secondary education −0.035 0.036 −0.045 0.036 −0.046 0.038
Treatment × higher education −0.071 0.059 −0.079 0.059 −0.074 0.064
Treatment × woman works agriculture −0.069 0.045 −0.067 0.045 −0.053 0.046
Treatment × woman is salaried worker 0.016 0.107 0.008 0.107 0.024 0.107
Treatment × woman has small business −0.015 0.075 −0.023 0.075 −0.001 0.075
Treatment × woman is a waged worker 0.005 0.071 0.021 0.071 0.012 0.073
Treatment × woman has other work −0.508** 0.238 −0.522** 0.243 −0.532** 0.234
Treatment × Terai and Madeshi 0.087* 0.052 0.105** 0.052 0.066 0.053
Treatment × Dalit 0.050 0.038 0.053 0.038 0.041 0.040
Treatment × Newar 0.146* 0.079 0.148* 0.079 0.123** 0.082
Treatment × Janajati 0.008 0.038 0.015 0.038 −0.012 0.043
Treatment × Muslim −0.044 0.184 −0.027 0.173 −0.087 0.179
Treatment × other castes 0.040 0.058 0.069 0.056 0.019 0.061
Treatment × time to facility 1 > 4 h −0.068* 0.041 −0.073* 0.041 −0.081 0.041
Treatment × time to facility 4 > 24 h −0.029 0.049 −0.025 0.050 −0.047 0.050
Treatment × time to facility > 1 day −0.048 0.053 −0.056 0.052 −0.071 0.056
Treatment × urban −0.035 0.060 −0.025 0.059 −0.056 0.060
Treatment × previous delivery SDIP 0.008 0.039 0.005 0.040 0.006 0.040
Treatment × active FCHV 0.067 0.102 0.054 0.104 0.070 0.102
Treatment × active women’s groups 0.011 0.028 0.006 0.028 0.012 0.029
Treatment × hill −0.103* 0.056 – – – –
Treatment × mountain −0.007 0.051 – – – –
Treatment × ratio of subsidy to cost – – 0.205** 0.096 0.996*** 0.310
Treatment × hospital quality – – – – 0.137** 0.059
Treatment × PHCC quality – – – – 0.085** 0.033
Treatment × health post quality – – – – −0.050 0.031
Treatment × year 0.008 0.018 0.011 0.018 0.011 0.018

R2 0.213 0.211 0.217

F-statistic for null that all interactions = 0 1.81 1.69 2.21
p-Value 0.0072 0.0171 0.0002
F-statistic for null that regional interactions = 0 4.32 – –
p-value 0.0134 – –
F-statistic for null that accessibility interactions = 0 1.21 1.51 1.50
p-Value 0.3046 0.209 0.212
F-statistic for null that caste interactions = 0 1.18 1.44 0.93
p-Value 0.3157 0.1936 0.4691

Note: Sample includes 5901 deliveries. Results are based on a linear probability model, which is weighted by the kernel weight provided by psmatch2. Each model includes all
the covariates that were interacted with the treatment indicator, but the effects of these variables are not reported. FCHV = female community health volunteer. PHCC = primary
health care centre.
*
Impact parameters are significant at 10%.
**
Impact parameters are significant at 5%.
***
Impact parameters are significant at 1%.

and the quality of care is low (Lagarde et al., 2007). We find that
the effect of the SDIP on utilisation was positively associated with
our proxy for quality of care at hospitals and primary health care
centres, which provides support to the argument that demand-side
incentives are most effective once supply-side improvements have
been put place. In the literature on conditional cash transfers, the
issue of interaction between supply and demand has been given
almost no attention and is a pressing area for further research, par-
ticularly in view of the fact that financial incentives are becoming
more widespread in low-income countries. The analysis of het-
erogeneity also finds that larger financial incentives are associated
with a greater impact on utilisation, as expected.
5.1. Interpretation of magnitudes
impact on utilisation if there were no problems in the disbursement
of cash to beneficiaries. This is useful in separating the question of
whether the incentive mechanism underpinning the programme
worked from questions about administrative failings. An alterna-
tive study design, such as randomisation of the programme across
intervention and control districts, would have struggled to detect
any impact because implementation was poor. The unusual study
design does mean, however, the findings should be accompanied by
information on implementation, particularly if the interpretation
is to be meaningful to policymakers.13
To further interpret the magnitudes of the effect of the SDIP
we take into account the extent of implementation. Because cover-
age of the SDIP was low, the increase in utilisation over the entire
sample is a fraction of the estimates given by the ATT. With just

The treatment group was defined in such a way that the measure
of programme impact essentially filters out problems in the imple-
mentation. The analysis of the SDIP’s promise of cash provides an
upper bound on the treatment effect and can be interpreted as the
13
Particular care should be taken in the interpretation of results for subgroups. It is
perfectly possible, and indeed likely, that the SDIP had a large impact on utilisation
of a subgroup (i.e. high ATT estimate), yet failed to reach many in that subgroup (i.e.
low uptake).
 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284 281

over one-quarter of the sample reached by the programme, the ATT Table A1
Probit model results for the estimation of the propensity score.
estimate implies that the increase in skilled birth attendance as a
result of the SDIP was 1.1 percentage points across the entire sam- Variable Coefficient Std error z-Stat
ple. Similarly, the rise in the utilisation of institutional deliveries Constant 7.029 5.420 1.30
attributable to the SDIP is 1.0 percentage points and, in the case of Age of woman −0.007 0.004 −1.94
the caesarean section rate, 0.3 percentage points. These effects are Log of wealth asset score −14.183 7.565 −1.87
modest, particularly when we consider the Government of Nepal’s Log of wealth asset score2 7.596 3.862 1.97
Log of wealth asset score3 −1.698 0.855 −1.99
target for skilled birth attendance is 60 percent in 2015.14
Log of wealth asset score4 0.138 0.069 1.99
There is a growing literature on demand-side incentives for Primary education 0.255 0.065 3.94
health against which to compare the magnitudes of our estimated Secondary education 0.294 0.061 4.78
effects, although few are specific to maternal health. Experimen- Higher education 0.529 0.083 6.40
Woman works in agriculture 0.134 0.062 2.18
tal evidence comes from studies of conditional cash transfers
Woman is salaried or government worker 0.265 0.145 1.82
in Mexico (Fernald et al., 2008; Gertler, 2000, 2004), Nicaragua Woman has small business 0.016 0.099 0.16
(Maluccio and Flores, 2005), Brazil (Morris et al., 2004b), Ecuador Woman is a waged worker 0.102 0.095 1.08
(Paxson and Schady, 2008) and Honduras (Morris et al., 2004a), Woman has other work −0.060 0.374 −0.16
one-off financial incentives in Malawi (Thornton, 2008), and Terai and Madeshi Other castes 0.042 0.083 0.51
Dalit −0.054 0.064 −0.84
non-financial incentives in India (Banerjee et al., 2010). The inter-
Newar −0.020 0.124 −0.16
ventions in these studies are targeted towards poor families and Janajati −0.198 0.061 −3.25
most provide some evidence of positive effects on utilisation of Muslim −0.639 0.206 −3.10
health services and immunization coverage.15 Other castes −0.142 0.081 −1.75
Walk to facility 1 h < 4 h −0.158 0.057 −2.76
In Malawi, a small-scale project was found to increase the per-
Walk to facility 4 h < 1 day −0.422 0.076 −5.52
centage of individuals who collected their HIV test results by 44 Walk to facility > 1 day −0.670 0.105 −6.36
percentage points (Thornton, 2008). Perhaps the most well-known Urban 0.125 0.078 1.59
CCT programme is Mexico’s Oportunidades, which was shown to Woman had a previous delivery during SDIP 0.130 0.074 1.76
increase health clinic consultations by 2.1 visits per day (Gertler, Active female community health volunteer 2.829 0.192 14.74
Active women’s groups 0.161 0.043 3.74
2000). The CCT programme in Honduras increased utilisation of
Sankhuwasabha 0.854 0.086 9.99
prenatal care by women, routine paediatric examinations and child Myagdi 0.600 0.083 7.25
growth monitoring by 19 percentage points, 20 percentage points Rupandehi 0.546 0.098 5.57
and 16 percentage points respectively (Morris et al., 2004a). A Jumla 0.020 0.074 0.27
Achham 0.651 0.104 6.25
similar programme in Nicaragua increased utilisation of child pre-
Year 0.336 0.023 14.54
ventive health visits by 11 percentage points (Maluccio and Flores, N 5905
2005). No significant impact of CCTs on health visits was found Pseudo R2 0.1765
in Ecuador (Paxson and Schady, 2008). Finally, in the Indian state Log likelihood −2745.6
of Rajasthan, lentils were offered alongside immunization camps,
raising full immunization rates to 39 percent compared with 6
percent in control and 18 percent in immunization camp only vil-
lages (Banerjee et al., 2010). The evidence on financial incentives expenditure on demand-side incentives amounts to US$ 115 for
in health is limited almost exclusively to the use of simple health each additional delivery in a health facility. When we factor in the
technologies. More complex health services, whose quality of care cost of delivery care using data from Borghi et al. (2006a), we esti-
is more difficult for patients to assess, have rarely been targeted mate the cost to be approximately US$ 210 per additional facility
using financial incentives. birth. This should be considered a lower bound estimate since we
While it is beyond the scope of this paper to conduct a detailed include no programme administration costs and we almost cer-
cost-effectiveness analysis,16 we are able to make a back-of-the- tainly have not captured the full economic cost of delivery care.
envelope calculation of the cost per additional facility birth.17 Using Given that the cash payment ranges from $8 in the tarai districts
information on programme spending, we estimate that the SDIP’s to around $40 in the mountain districts, the estimate of US$ 115
gives some sense of the inefficiency of providing universal financial
incentives.
The combination of the two sets of findings on impact and
14
According to our data, coverage of skilled birth attendance grew from 19 percent implementation suggest a missed opportunity for the govern-
to 24 percent over the period 2005–2008, still well below the 60 percent government
ment. Because implementation was far from complete, the financial
target.
15
We report below effects that were shown to be significant but note that the incentives were unable to bite and in practice few women were
studies also show evidence of no effect on numerous other utilisation outcomes, incentivised to seek formal care at childbirth.18 These results imply
which are summarized in a systematic review of CCT programmes by Lagarde et al. that the impact of the SDIP on utilisation over the entire popula-
(2007). Robust estimates from nonexperimental studies include CCT programmes
tion has been too low to generate concerns of maternity services
in Columbia (Attanasio et al., 2005), Turkey (Ahmed et al., 2007), and Chile (Galasso,
2007).
being overwhelmed from an influx of deliveries. Indeed, there
16
In particular, we consider it too speculative to attempt to impute an estimate of was no evidence from the qualitative studies in the evaluation
lives saved from our utilisation impact results given that there is little or no rigorous to suggest otherwise. The magnitude of the impact estimates also
evidence on the effect of facility births on health outcomes (Campbell and Graham, suggest there can be little expectation of a sizeable improvement
2006).
17
The cost is based on SDIP expenditure on the demand-side incentives over the
period 2005–2008 and estimates of the cost of giving birth in a health facility from
Borghi et al. (2006a), adjusted for inflation. We use an exchange rate of 70 Nepalese
Rupees per dollar. The cost does not include the cost of SDIP administration, nor population projections and estimates of the crude birth rate from the Census 2001
the cost of technical assistance provided to the programme by the Support to Safe and the Demographic and Health Survey 2006 respectively.
18
Motherhood Programme. The effect of the SDIP is calculated on the basis of the We can speculate further that public’s experience of the programme may have
1.0 percentage point increase in the number of institutional deliveries attributable eroded the effectiveness of future policy by undermining trust in the government
to the demand-side incentives. The total number of deliveries is calculated using to meet its stated commitments.
282 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284

Table A2
Balancing of the covariates by type of matching estimator.

Variable Before matching After kernel matching After mahalanobis matching After nearest neighbour matching

Bias (%) t-Stat Bias (%) t-Stat % Reduction Bias (%) t-Stat % Reduction Bias (%) t-Stat % Reduction
in bias in bias in bias

Age of woman −22.0 −7.13 −1.2 −0.33 94.6 1.6 0.47 92.8 −1.7 −0.48 92.3
Ln(wealth asset score) 22.7 7.52 2.3 0.64 89.8 3.2 0.90 85.7 3.0 0.83 86.9
Ln(wealth asset score)2 21.2 7.08 2.4 0.66 88.6 3.5 0.96 83.4 3.2 0.87 84.9
Ln(wealth asset score)3 19.7 6.67 2.4 0.65 87.6 3.7 0.98 81.5 3.3 0.89 83.1
Ln(wealth asset score)4 18.5 6.31 2.4 0.63 86.9 3.7 0.98 79.9 3.4 0.89 81.6
Primary education 4.9 1.66 −1.2 −0.32 75.2 3.8 1.04 21.7 −2.6 −0.67 47.5
Secondary education 18.1 6.22 −1.7 −0.44 90.6 −1.4 −0.37 92.1 −1.3 −0.33 93.0
Higher education 32.2 11.83 4.9 1.17 84.8 3.8 0.90 88.2 4.3 1.03 86.7
Woman works in agriculture −3.8 −1.26 −0.3 −0.09 90.7 −3.6 −0.99 3.2 −0.5 −0.12 87.9
Woman is salaried worker 12.3 4.61 3.1 0.75 74.5 0.0 0.00 100.0 1.8 0.42 85.5
Woman has small business 8.6 3.01 −3.5 −0.85 59.7 1.2 0.30 86.5 −4.6 −1.10 47.3
Woman is a waged worker −8.1 −2.62 −2.8 −0.79 65.6 1.8 0.54 77.7 −0.4 −0.12 94.8
Woman has other work 2.5 0.91 1.0 0.25 61.0 0.0 0.00 100.0 1.0 0.26 58.9
Terai and Madeshi other castes −6.3 −2.04 −1.3 −0.38 78.9 1.0 0.30 83.4 −2.9 −0.81 53.5
Dalit −6.9 −2.26 0.6 0.16 92.0 1.6 0.46 76.8 1.7 0.48 75.7
Newar 16.1 6.10 1.0 0.22 94.1 0.0 0.00 100.0 0.2 0.05 98.7
Janajati −7.9 −2.59 0.2 0.06 97.3 −0.7 −0.19 91.3 0.5 0.14 93.7
Muslim −18.4 −5.28 −0.2 −0.11 98.8 0.0 0.00 100.0 0.1 0.03 99.7
Other castes −18.0 −5.69 −1.1 −0.33 94.1 0.2 0.07 98.8 −0.9 −0.28 94.9
Distance to facility 1 > 4 h 8.4 2.81 −0.2 −0.04 98.2 −6.7 −1.86 20.7 0.7 0.20 91.2
Distance to facility 4 > 24 h −19.6 −6.24 −0.3 −0.10 98.3 2.3 0.70 88.4 −0.8 −0.24 96.0
Distance to facility > 1 day −29.9 −8.89 −0.3 −0.12 98.9 0.5 0.21 98.2 −0.8 −0.32 97.2
Urban 12.1 4.22 3.2 0.83 73.6 5.1 1.33 57.8 3.6 0.93 70.2
Previous delivery during SDIP 11.5 4.04 1.5 0.37 87.2 11.7 3.20 −1.6 −1.0 −0.26 91.1
Active FCHV 62.1 22.20 −2.9 −0.69 95.3 12.8 3.18 79.3 −4.3 −1.01 93.2
Active women’s groups 13.8 4.66 3.6 0.96 73.9 9.5 2.57 31.0 2.7 0.70 80.7
Sankhuwasabha 25.1 8.83 −2.0 −0.51 91.8 1.0 0.26 95.9 −1.8 −0.44 92.9
Myagdi 5.8 1.95 1.3 0.36 76.6 1.3 0.35 77.3 1.0 0.26 83.0
Rupandehi 14.7 5.05 −1.9 −0.48 87.3 0.3 0.09 97.7 −2.6 −0.65 82.6
Jumla −21.5 −6.82 0.3 0.08 98.8 −1.1 −0.35 94.7 0.8 0.25 96.2
Achham −16.5 −5.29 3.0 0.92 81.5 −0.4 −0.11 97.8 2.8 0.85 82.9
Year 40.5 13.06 −1.4 −0.38 96.6 22.8 6.64 43.7 −0.1 −0.02 99.9

in maternal health, particularly when one accepts that increased
utilisation does not automatically translate into improved health.
5.2. Limitations
Discussion of a number of issues allows consideration of the
limitations of the analysis. First, the findings rest on the plausibility
of the conditional independence assumption, which is ultimately
untestable. We assessed the robustness of the findings to this
assumption and on the basis of these results believe we are able to
make causal inferences from our findings. In further support of this
conclusion, we would argue that internal validity is strengthened
by the richness of the data (Heckman et al., 1998; Diaz and Handa,
2006) and the fact that selection bias may be less of a problem in
the context of this study because individuals do not actively select
or enrol into the programme. Rather, selection into treatment is a
passive process and it seems reasonable to assume that selection
bias on unobservables is more of a problem in programmes which
require a decision on the part of the beneficiary to participate, since
the decision may be influenced by, rather than simply correlated
with, the unobservables.
Second, the sample was not designed to be representative of
the entire country, which means the results should not be inter-
preted as national impact estimates. Third, our study covered only
the early period of the SDIP and all indications are that the pro-
gramme has improved since our household survey. Finally, our
analysis was limited to utilisation outcomes as the main measure
of performance. In contrast with many simple health technologies,
the link between utilisation of maternity services and improved
health depends on a complex set of factors linked to the qual-
ity of care. Individuals demand health, not health care per se.
Judging the success of the SDIP on the extent to which it influenced
health seeking behaviour at childbirth is thus fraught with prob-
lems. The standard measure of utilisation, skilled birth attendance,
is a crude measure of access to quality maternity services. Not only
does it fail to capture well the skills of the attendant, it says noth-
ing about many other important aspects of quality that impact on
health. Ideally, we would have measured maternal deaths as the
main outcome of interest but because they are such a rare event
the sample size would have been prohibitively large to detect even
a large effect.19
6. Conclusion
In this paper we find that a national programme offering
financial incentives to households was modestly effective in
encouraging women to deliver with professional care in six dis-
tricts of Nepal. The impacts appear to be modified by the size of
the financial package relative to the cost of care and the quality of
care provided in hospitals and primary health care centres. Owing
to the low coverage of the SDIP, a small proportion of the popu-
lation were incentivised by the programme to seek formal care at
childbirth, illustrating how slow implementation has constrained
its success.
19
This point is well illustrated by a randomised experiment of vitamin A sup-
plementation on maternal survival in Ghana which lasted seven years and required
over 200,000 participants to conclude that the intervention had no effect (Kirkwood
et al., 2010).
 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284 283

Table A3
Impact estimates by matching procedure.

Variable Kernel Mahalanobis Nearest neighbour

ATT t-Stat ATT t-Stat ATT t-Stat

Place of delivery
Health facility 0.040*** 2.70 0.039** 2.06 0.046*** 3.04
Government health facility 0.043*** 3.28 0.045** 2.65 0.048*** 3.52
NGO hospital −0.011* −1.86 −0.009 −1.17 −0.011* −1.73
Private health facility 0.008 1.40 0.003 0.48 0.009 1.57

Type of attendant
Doctor, nurse or midwife (SBA) 0.042*** 2.72 0.042** 2.11 0.048*** 3.00
Any professional health worker 0.052*** 3.17 0.045** 2.16 0.058*** 3.43

Procedure at delivery
Caesarean section 0.012* 1.83 0.018** 2.13 0.013* 1.83
Caesarean section or assisted 0.019** 2.02 0.032*** 2.67 0.022** 2.25

Note: NGO = non-governmental organisation; SBA = skilled birth attendant.

*
Significance at 10%.
**
Significance at 5%.
***
Significance at 1%.

Acknowledgements
We have benefited enormously from discussions with Anthony
Costello, Anne Mills, Suresh Tiwari, Basu Dev Neupane, Louise
Hulton, Greg Whiteside and seminar participants at iHEA, Oxford
University and University College London. We thank the dedicated
group of individuals involved in the data collection. The study was
funded by the Department for International Development through
the Support to Safe Motherhood Programme in Nepal (managed by
Options UK). Financial support from the ESRC and MRC is gratefully
acknowledged. The research was approved by the ethics committee
of the London School of Hygiene and Tropical Medicine. All errors
are our own.
Appendix A.
See Tables A1–A3.
References
Ahmed, A., Adato, M., Kudat, A., Gilligan, D., Roopnaraine, T., Colasan, R., 2007. Impact
Evaluation of the Conditional Cash Transfer Program in Turkey: Final Report.
International Food Policy Research Institute, Washington, DC.
Angrist, A.A., Pischke, J., 2009. Mostly Harmless Econometrics: An Empiricist’s Com-
panions. Princeton University Press, Princeton.
Attanasio, O., Gómez, L., Heredia, P., Vera-Hernández, M., 2005. The Short-term
Impact of a Conditional Cash Subsidy on Child Health and Nutrition in Colombia.
Institute of Fiscal Studies, London.
Banerjee, A.V., Duflo, E., Glennerster, R., Kothari, D., 2010. Improving immunisation
coverage in rural India: clustered randomised controlled evaluation of immu-
nisation campaigns with and without incentives. BMJ 340, c2220.
Borghi, J., Ensor, T., Neupane, B.D., Tiwari, S., 2006a. Financial implications of skilled
attendance at delivery in Nepal. Tropical Medicine and International Health 11,
228–237.
Borghi, J., Ensor, T., Somanathan, A., Lissner, C., Mills, A., 2006b. Mobilising financial
resources for maternal health. Lancet 368, 1457–1465.
Campbell, O.M., Graham, W.J., 2006. Strategies for reducing maternal mortality:
getting on with what works. Lancet 368, 1284–1299.
Cochran, W.G., Rubin, D.B., 1973. Controlling bias in observational studies. Sankhya
– The Indian Journal of Statistics Series A 35, 417–446.
Diaz, J.J., Handa, S., 2006. An assessment of propensity score matching as a non-
experimental impact estimator—Evidence from Mexico’s PROGRESA program.
Journal of Human Resources 41, 319–345.
Djebbari, H., Smith, J., 2008. Heterogeneous impacts in PROGRESA. Journal of Econo-
metrics 145, 64–80.
Ensor, T., Clapham, S., Prasai, D.P., 2009. What drives health policy formulation:
insights from the Nepal maternity incentive scheme? Health Policy 90, 247–253.
Fernald, L.C., Gertler, P.J., Neufeld, L.M., 2008. Role of cash in conditional cash transfer
programmes for child health, growth, and development: an analysis of Mexico’s
Oportunidades. Lancet 371, 828–837.
Filmer, D., Pritchett, L., 1998. Estimating Wealth Effects with No Expenditure Data
– Or Tears: An Application to Educational Enrollment in States of India. World
Bank, Washington, DC.
Fiszbein, A., Schady, N., Ferreira, F.H., Grosh, M., Kelleher, N., Olinto, P., Skofias,
E., 2008. Conditional Cash Transfers: Reducing Present and Future Povertys.
IBRD/The World Bank.
Galasso, E., 2007. With their Effort and One Opportunity: Alleviating Extreme
Poverty in Chile. World Bank, Washington, DC.
Gertler, P., 2000. Final Report: The Impact of PROGESA on Health. International Food
Policy Research Institute, Washington, DC.
Gertler, P., 2004. Do conditional cash transfers improve child health? Evidence from
PROGRESA’s control randomized experiment. American Economic Review 94,
336–341.
Gertler, P., Locay, L., Sanderson, W., 1987. Are user fees regressive? The welfare
implications of charging for medical care in Peru. Journal of Econometrics 36
(supp), 67–68.
Gertler, P., Van der Gaag, J., 1990. The Willingness to Pay for Medical Care: Evidence
from Two Developing Countries. Johns Hopkins University Press, Baltimore and
London.
Government of Nepal, 2001. Nepal Demographic and Health Surveys. ORC Macro,
MD, USA/Kathmandu, Nepal.
Government of Nepal, 2005. Operational guidelines on incentives for safe delivery
services. Government of Nepal, Ministry of Health and Population, Kathmandu.
Government of Nepal, 2007. Nepal Demographic and Health Survey. ORC Macro,
MD, USA/Kathmandu, Nepal.
Heckman, J., Ichimura, H., Todd, P., 1998. Matching as an econometric evaluation
estimator. Review of Economic Studies, 261–294.
Imbens, G.W., Angrist, J.D., 1994. Identification and estimation of local average treat-
ment effects. Econometrica 62, 467–475.
Imbens, G.W., Wooldridge, J.M., 2009. Recent Developments in the Econometrics of
Program Evaluation. Journal of Economic Literature 47, 5–86.
Jack, A., 2008. World Bank backs anti-Aids experiment. Financial Times, London.
Kirkwood, B.R., Hurt, L., Amenga-Etego, S., Tawiah, C., Zandoh, C., Danso, S., Hurt,
C., Edmond, K., Hill, Z., Ten Asbroek, G., Fenty, J., Owusu-Agyei, S., Campbell, O.,
Arthur, P., 2010. Effect of vitamin A supplementation in women of reproductive
age on maternal survival in Ghana (ObaapaVitA): a cluster-randomised, placebo-
controlled trial. Lancet 375, 1640–1649.
Koblinsky, M., Matthews, Z., Hussein, J., Mavalankar, D., Mridha, M.K., Anwar, I.,
Achadi, E., Adjei, S., Padmanabhan, P., Marchal, B., De Brouwere, V., van Ler-
berghe, W., 2006. Going to scale with professional skilled care. Lancet 368,
1377–1386.
Lagarde, M., Haines, A., Palmer, N., 2007. Conditional cash transfers for improving
uptake of health interventions in low- and middle-income countries: a system-
atic review. JAMA 298, 1900–1910.
Leuven, E., Sianesi, B., 2003. PSMATCH2: Stata module to perform full Mahalanobis
and propensity score matching, common support graphing, and covariate imbal-
ance testings.
Maluccio, J., Flores, R., 2005. Impact Evaluation of a Conditional Cash Transfer
Program: The Nicaraguan Red de Protección Social. International Food Policy
Research Insitute, Washington, DC.
Morris, S., Flores, R., Olinto, P., Medina, J.M., 2004a. Monetary incentives in primary
health care and effects on use and coverage of preventive health care interven-
tions in rural Honduras: cluster randomised trial. Lancet 364, 2030–2037.
Morris, S.S., Olinto, P., Flores, R., Nilson, E.A., Figueiro, A.C., 2004b. Conditional
cash transfers are associated with a small reduction in the rate of weight
gain of preschool children in northeast Brazil. Journal of Nutrition 134,
2336–2341.
NORA.D, 2007. The Global Campaign for the Health Millennium Development
Goalss. NORAD, Oslo.
Oxman, A., Fretheim, A., 2008. An Overview of Research on the Effects of
Results-based Financing vol. Report Nr 16-2008. Nasjonalt kunnskapssenter for
helsetjenest, Oslo.
284 T. Powell-Jackson, K. Hanson / Journal of Health Economics 31 (2012) 271–284
Paxson, C., Schady, N., 2008. Does Money Matter? The Effects of Cash Transfers on
Child Health and Development in Rural Ecuador. World Bank, Washington, DC.
Powell-Jackson, T., Morrison, J., Tiwari, S., Neupane, B.D., Costello, A.M., 2009a. The
experiences of districts in implementing a national incentive programme to
promote safe delivery in Nepal. BMC Health Servies Research 9, 97.
Powell-Jackson, T., Neupane, B., Tiwari, S., Tumbahangphe, K., Manandhar, D.,
Costello, A., 2009b. The impact of Nepal’s national incentive progamme to pro-
mote safe delivery in the district of Makwanpur. In: Chernichovsky, D., Hanson,
K. (Eds.), Innovations in Health System Finance in Developing and Transitional
Economies. Emerald, Bingley.
Pradhan, A., Aryal, R., Regmi, G., Ban, B., Govindasamy, P., 1997. Nepal Family Health
Survey 1996. Government of Nepal, Ministry of Health/Macro International Inc.,
New Era/Calverton, Maryland and Kathmandu.
Rohde, J., Cousens, S., Chopra, M., Tangcharoensathien, V., Black, R., Bhutta, Z.A.,
Lawn, J.E., 2008. 30 years after Alma-Ata: has primary health care worked in
countries? Lancet 372, 950–961.
Rosenbaum, P., Rubin, D., 1983. The central role of the propensity score in observa-
tional studies for causal effects. Biometrika 70, 41.
Rosenbaum, P., Rubin, D., 1984. Reducing bias in observational studies using
subclassification on the propensity score. Journal of the American Statistical
Association, 516–524.
Rosenbaum, P., Rubin, D., 1985. Constructing a control group using multivariate
matched sampling methods that incorporate the propensity score. The Ameri-
can Statistician 39, 33–38.
Sargan, J., 1958. The estimation of economic relationships using instrumental vari-
ables. Econometrica 26, 393–415.
Smith, J.A., Todd, P.E., 2005. Does matching overcome LaLonde’s critique of nonex-
perimental estimators? Journal of Econometrics 125, 305–353.
Thornton, R.L., 2008. The demand for, and impact of, learning HIV status. American
Economic Review 98, 1829–1863.
UNICEF, 1997. Guidelines for Monitoring the Availability and Use of Obstetric Ser-
vicess. UNICEF, New York.
Wagstaff, A., Lindelow, M., Jun, G., Ling, X., Juncheng, Q., 2009. Extend-
ing health insurance to the rural population: an impact evaluation of
China’s new cooperative medical scheme. Journal of Health Economics 28,
1–19.
Weiss, C., 1998. Evaluation: Methods for studying programs and policies. Prentice
Hall.
White, H., Masset, E., 2007. Assessing interventions to improve child nutrition: a
theory-based impact evaluation of the Bangladesh Integrated Nutrition Project.
Journal of International Development 19, 627.